Commit Graph

  • 339f3f10a3 Use HND layout leo/prefill-decode-really Ryuichi Leo Takashige 2026-04-01 19:58:12 +01:00
  • 37ad1fb3ed Call warmup_speculative at startup to pre-compile LpB kernels david/batched-kernels dmcc73 2026-04-01 23:15:05 +01:00
  • 42ad50991e runner failling should imply the runner is dead failed=stopped Evan 2026-04-01 15:29:31 +01:00
  • eb6ae9fd3c Prevent failed instance retries (#1763) main ciaranbor 2026-04-01 21:03:34 +01:00
  • f5115827ce Monitor instance health Don't checkpoint failed evals Auto-delete stale instances before placement ciaran/eval ciaranbor 2026-03-31 14:59:21 +01:00
  • 8290afa642 Don't timout for downloads ciaranbor 2026-03-31 14:59:19 +01:00
  • 19adb4a763 Add models ciaranbor 2026-03-31 14:58:46 +01:00
  • b7b89f6e0f Remove power_usage from response in favour of existing energy polling in exo bench ciaranbor 2026-03-31 14:15:52 +01:00
  • 7339b6426d Fix Kimi-K2.5 tokenization ciaranbor 2026-03-19 16:21:05 +00:00
  • 2555b0e649 Resolve higher-node placements first ciaranbor 2026-03-18 15:10:03 +00:00
  • a4101918bf Add power/energy measurements to exo_bench ciaranbor 2026-03-17 17:18:04 +00:00
  • 58f5046de1 Extend eval harness: - top_k, min_p, enable_thinking params - capture reasoning_content, finish_reason, power/energy - checkpoint/resume support - instance resuse - LCB release_version param ciaranbor 2026-03-16 19:55:06 +00:00
  • 4f623f4090 Align models.toml configs with vllm run ciaranbor 2026-03-16 19:54:50 +00:00
  • bed9d9220d Update download progress on pause download-pause-delete ciaranbor 2026-04-01 19:52:54 +01:00
  • 5d22805a77 Benchmarking Ryuichi Leo Takashige 2026-04-01 19:22:45 +01:00
  • 68a4cf728d Fix rebase ciaran/storage-management ciaranbor 2026-04-01 11:41:17 +01:00
  • e6bee75d91 Use slider for storage limit ciaranbor 2026-03-20 13:17:33 +00:00
  • 70dbc03b37 Remove ModelEvicted status ciaranbor 2026-03-16 19:01:44 +00:00
  • b54cd7415d Rename model download statuses ciaranbor 2026-03-16 16:01:52 +00:00
  • fdd6d94fcc Add storage config IO tests ciaranbor 2026-03-12 17:57:19 +00:00
  • 26aacd55b7 Batch update storage configs ciaranbor 2026-03-12 12:18:35 +00:00
  • fee0d3fc48 Better separation of concerns ciaranbor 2026-03-12 11:40:08 +00:00
  • 0a951737ee Move rejection-triggered instance cleanup from _plan to _event_processor ciaranbor 2026-03-10 12:32:13 +00:00
  • a0fad061e2 Include ongoing downloads in calculate_used_storage ciaranbor 2026-03-10 11:27:56 +00:00
  • 9832f161e7 Evict directly instead of through DownloadRejected ciaranbor 2026-03-10 10:40:22 +00:00
  • e1ddaf97bd Pydantic <-> toml ciaranbor 2026-03-09 20:27:27 +00:00
  • fa29d0fd00 Deduplicate by model ID ciaranbor 2026-03-09 18:59:47 +00:00
  • 60f6d8e6fc Use DownloadEvicted event ciaranbor 2026-03-09 16:56:32 +00:00
  • f87d705b9c Implement storage management ciaranbor 2026-03-06 19:00:27 +00:00
  • 94a7264cbf Add cli arguments to set max storage and eviction policy ciaranbor 2026-03-06 18:30:05 +00:00
  • bf35249d9a Add StorageConfig to state, with max storage and storage policy settings ciaranbor 2026-03-06 18:26:36 +00:00
  • 9d5242a25e Add StorageConfigUpdated event ciaranbor 2026-03-06 18:24:39 +00:00
  • f88aea5bbe Add SetStorageConfig command ciaranbor 2026-03-06 18:24:24 +00:00
  • 922e00ddc8 Add DownloadRejected type ciaranbor 2026-03-06 18:24:14 +00:00
  • d0d71ebdb2 Cancel SSE keep-alive when instance is deleted streaming-delete-hang ciaranbor 2026-04-01 17:49:14 +01:00
  • 89750121f4 Allow pausing and deleting active downloads ciaranbor 2026-04-01 18:38:52 +01:00
  • 03c7b7627d Allow small num_heads leo/add-uneven-sharding Ryuichi Leo Takashige 2026-04-01 12:26:10 +01:00
  • e78454de41 Don't reject uneven placements in placement Ryuichi Leo Takashige 2026-03-31 20:13:12 +01:00
  • b30ee156c4 Add advanced dashboard elements for tensor sharding and update exo bench Ryuichi Leo Takashige 2026-03-31 19:22:17 +01:00
  • ae7ba5d054 Messy POC Ryuichi Leo Takashige 2026-03-31 17:30:56 +01:00
  • 55fa5362bb Add uneven sharding Ryuichi Leo Takashige 2026-03-31 13:22:05 +01:00
  • 4688adb5d2 Support PDFs in dashboard (#1822) simplify-downloads rltakashige 2026-03-31 18:25:40 +01:00
  • 499edece9b bump versions correctly vllm-nix Evan 2026-03-31 12:45:55 +01:00
  • 784662f40a MTP speculative decoding + kernel patches for Qwen3.5 (rebased on main) david/speculative-mtp dmcc73 2026-03-31 00:04:54 +01:00
  • b47a287f3e Add EXO_DISABLE_LOGPROBS=1 to skip per-token logprobs extraction dmcc73 2026-03-30 23:41:47 +01:00
  • e1cf376e45 Add speculative warmup: compile MTP + verify kernels at startup dmcc73 2026-03-30 22:39:34 +01:00
  • 75932cbcca Fix stop token dropping valid tokens before it dmcc73 2026-03-30 18:44:01 +01:00
  • 199a4ab7e0 EXO_SPECULATIVE_TEMP overrides model sampling temperature globally dmcc73 2026-03-30 18:24:25 +01:00
  • 4818b9a3db EXO_SPECULATIVE_TEMP overrides request temp when set dmcc73 2026-03-30 18:23:05 +01:00
  • f0433505a8 Fix speculative temp: use request temperature, not global env var dmcc73 2026-03-30 18:21:57 +01:00
  • 8a2d05580f allow flexible apple-sdk Evan 2026-03-30 18:01:51 +01:00
  • 88bc1656a2 Fix MTP prefill to use all captured positions dmcc73 2026-03-30 18:06:53 +01:00
  • d9ed943034 Fix Nemotron cache leak upstream (#1819) rltakashige 2026-03-30 17:53:21 +01:00
  • 7bd1ba6605 Fix MTP prefill: do it in submit() with correct prompt tokens dmcc73 2026-03-30 17:42:20 +01:00
  • e7c5d56e83 Add LpB kernel patches for Qwen3.5 dense models (27B, 9B) dmcc73 2026-03-30 17:00:02 +01:00
  • dd71182457 Fix MTP prefill for exo: capture prompt tokens via embed_tokens wrapper dmcc73 2026-03-30 15:57:25 +01:00
  • 436a7e3dcd fix mac build Evan 2026-03-30 15:42:13 +01:00
  • 09012d3799 Auto-extract MTP weights from HuggingFace model repo dmcc73 2026-03-30 15:19:48 +01:00
  • ce19267d2d Pass temperature and alpha to MTP speculative decoding dmcc73 2026-03-30 15:12:59 +01:00
  • 8a65a51569 Add MTP speculative decoding for Qwen3.5 models dmcc73 2026-03-30 15:10:11 +01:00
  • c6815bfdce Only update KV prefix cache on a good cache hit (#1817) rltakashige 2026-03-30 15:04:38 +01:00
  • 39c39e8199 Integrations helpers (#1810) rltakashige 2026-03-30 14:28:41 +01:00
  • e5cb7b80d0 Add SSE-keepalive to not time out on long prefill on clients (#1803) rltakashige 2026-03-30 12:18:38 +01:00
  • 4166c04f6b build mlx cpu for now Evan 2026-03-30 12:07:17 +01:00
  • a1618c7f98 ITS ALIVE Evan 2026-03-30 12:07:16 +01:00
  • 35f57c2d3c think thats it Evan 2026-03-30 12:07:15 +01:00
  • 8753354d0c run the formatter leo Evan 2026-03-30 12:07:14 +01:00
  • 635801d515 Add multimodality! (#1802) rltakashige 2026-03-30 11:52:19 +01:00
  • 2efbb8ab4f Improve exo harness with path state (#1815) rltakashige 2026-03-30 11:20:46 +01:00
  • c6c5a3e73c feat: /state/paths (#1796) Evan Quiney 2026-03-30 11:10:00 +01:00
  • 10ef7ec9e8 feat: add Firefox AI sidebar (?q=) support to dashboard (#1814) ArvidSU 2026-03-30 12:02:35 +02:00
  • 9b03b561d3 Tighten up timings a bit more leo/exo-bench-conc-fix Ryuichi Leo Takashige 2026-03-27 23:46:00 +00:00
  • e9d911ccfc Add both stats Ryuichi Leo Takashige 2026-03-27 21:23:37 +00:00
  • 5bf5645b20 Send at the same time to improve exo bench concurrency accuracy Ryuichi Leo Takashige 2026-03-27 21:15:37 +00:00
  • fd17a3bc09 vllm loads! runner-refactor rltakashige 2026-03-27 12:02:36 +00:00
  • 1e51dc89b0 chore: bump exo-version with release version (#1807) Evan Quiney 2026-03-27 11:47:13 +00:00
  • 5327bdde84 Fix custom model add requiring two attempts + enlarge sidebar buttons (#1805) v1.0.69 Alex Cheema 2026-03-26 17:57:00 -07:00
  • 15f1b61f4c Rework model storage directory management (for external storage) (#1765) ciaranbor 2026-03-26 17:46:46 +00:00
  • 9034300163 [Fix] Node hang on reelection (#1801) Michael Harrigan 2026-03-26 13:28:47 -04:00
  • 1d1dfaa1f3 Don't download original/ and metal/ folders from HF (#1800) rltakashige 2026-03-26 13:36:07 +00:00
  • 7625213df0 fix: enable macmon if preflight fails (#1799) Evan Quiney 2026-03-26 11:35:22 +00:00
  • f318f9ea14 Fix macOS build bundling wrong macmon binary (#1797) Alex Cheema 2026-03-26 02:47:48 -07:00
  • 30fd5aa1cc Prefer higher % downloaded nodes for API placement previews (#1795) ciaranbor 2026-03-25 17:26:31 +00:00
  • 6de14cfedb Support image generation cancellation (#1774) ciaranbor 2026-03-25 16:56:04 +00:00
  • c6fab7fa97 fix: only shutdown runners that error unrecoverably runner-shutdown-vs-failed Evan 2026-03-25 15:39:14 +00:00
  • fc1ae90111 fix: DeepSeek V3.2 warmup crash and tool calling + add catalog cards (#1769) vskiwi 2026-03-25 19:20:35 +03:00
  • 565ed41c13 Fix occasional warmup bugs by using mlx_generate (#1794) rltakashige 2026-03-25 15:53:54 +00:00
  • 2da740c387 Feat/static peer discovery (#1690) DeepZima 2026-03-25 11:55:12 +01:00
  • 7ca9edaf42 wohohoh rltakashige 2026-03-25 10:35:46 +00:00
  • 7117d748ec Update dependencies including mlx 0.31.2 (#1789) rltakashige 2026-03-25 06:03:19 +00:00
  • 178c617bbb Rename Nemotron to NVIDIA in model picker with logo (#1790) Alex Cheema 2026-03-24 18:56:59 -07:00
  • 781428a176 Fix local model launch preparation handling alexcheema/fix-instance-preparing-delay Alex Cheema 2026-03-24 16:54:01 -07:00
  • 7277c90389 Fix enable thinking (#1786) rltakashige 2026-03-24 23:26:25 +00:00
  • 7ee88c1f05 override macmon in flake (#1747) Evan Quiney 2026-03-24 17:30:19 +00:00
  • 509533d49e send error finish reason on failing to parse a tool call (#1785) Evan Quiney 2026-03-24 17:21:39 +00:00
  • b6240a97e8 Prevent Qwen3.5 looping by using mlx lm fork (#1784) rltakashige 2026-03-24 17:16:41 +00:00
  • 6cdfbb7e8b Add HF_ENDPOINT in the app settings (#1783) rltakashige 2026-03-24 17:05:49 +00:00
  • fac6832e5f fix warmup consistency for slow machines (#1748) Evan Quiney 2026-03-24 16:45:55 +00:00
  • 7df3774ca2 Improve batch performance and stats reporting (#1777) rltakashige 2026-03-24 14:03:03 +00:00
  • 248919c2a8 Fix first start in offline mode crash (#1782) ciaranbor 2026-03-24 13:58:23 +00:00